Multiple Kernel Learning Improved By Bi-objective Functions And Its Application To Semi-supervised Learning And Transfer Learning

In recent year, kernel trick has become an important research area in data mining and machine learning. It utilizes the kernel function to build the similarity between the examples, and then construct the feature space for all data. As the latest kernel method, kernel learning has been paid more and more attentions recently. Different from traditional kernel methods, the target of kernel learning is to learn a suitable kernel function dynamically, by utilizing the maximum margin or the alignment, which will produces a better feature space. For the dynamic of kernel learning, it performs better than the traditional kernel methods. However, there are some definite shortages for kernel learning, firstly, the efficiency of kernel learning is low; secondly, the learning scenario of kernel learning is mostly assumed to be the supervised learning scenario. The researches on semi-supervised learning and transfer learning are focus on how to deal with the scenario that the training set is limited. The algorithm of kernel learning may not perform well in such scenarios for its little consideration for semi-supervised learning and transfer learning.To deal with the problems mentioned above, this paper proposes an multiple kernel learning (MKL) framework improved by bi-objective functions, which is based on the state of the art on kernel learning and has high efficiency. In this framework, traditional multiple kernel learning can combine with suitable objective functions and deal with the specific learning scenario. For the semi-supervised learning and transfer learning, this paper proposes the MKL improved by graph-laplace and the MKL improved by Maximum Mean Discrepancy(MMD) to deal with such scenarios. This paper also makes research on parameter between the objective functions to deal with the balance problem for bi-objective functions framework. The experiment of this paper is based on the real world datasets, the experiment results show that the MKL improved by graph-laplace and the MKL improved by MMD perform better than the traditional kernel methods in semi-supervised learning scenario and transfer learning scenario, and have comparable performances with the specific algorithm for such scenarios.